Institutional 13F
The 45-day deadline is the smaller half of the lag, and the number everyone quotes. A 13F is a snapshot of a quarter end, filed up to 45 days later. A position opened on the first day of that quarter is already 91 days old when the quarter closes. ✅ Measured below, the median position is 82 calendar days old at the moment its table becomes public, and the p95 is 127.
Everything marked ✅ Measured is printed by scripts/thirteen_f.py (numpy + pandas, seed
20260910, about 15 s, no network). Its holdings are synthetic and seeded. Everything
marked ✅ source-verified was read at the URL given, on 2026-09-10.
1. ✅ The rule, at primary sources
| Fact | Value | Source, read 2026-09-10 |
|---|---|---|
| Deadline | "within 45 days after the last day of such calendar year" and "within 45 days after the last day of each of the first three calendar quarters" | 17 CFR 240.13f-1(a)(1) (eCFR); Form 13F General Instruction 1 |
| Threshold | accounts holding section 13(f) securities "having an aggregate fair market value on the last trading day of any month" of "at least $100,000,000" | 17 CFR 240.13f-1(a)(1) |
| 🚨 Threshold never indexed | the rule's own source credit starts 43 FR 26705, June 22, 1978 and the figure survives every amendment. The SEC's 2020 proposal said it would be raising it "for the first time in 45 years, ... from $100 million to $3.5 billion" | Rel. 34-89290 (govinfo.gov) |
| ⚠️ The proposal was withdrawn | 2021-05-11, RIN 3235-AM65, "This item is being withdrawn." The $100M threshold stands. The SEC's own rule page still shows the proposal with no withdrawal note. | OMB/OIRA Unified Agenda |
| Format | "On May 20, 2013, the text-based ASCII format for 13F filings was discontinued." XML information table since. | 13F FAQ |
2. 🚨 What is in a 13F, and what is not
| instrument | status | source, read 2026-09-10 |
|---|---|---|
| US exchange-traded stock | reported | FAQ Q7: the Official List "primarily includes U.S. exchange-traded stocks" |
| closed-end funds, ETFs | reported | FAQ Q7 names both |
| convertible debt, equity options, warrants | reported if on the Official List | FAQ Q7: "Certain convertible debt securities, equity options, and warrants are on the Official List" |
| 🚨 short positions | NOT reported | FAQ Q41 verbatim: "You should not include short positions on Form 13F." |
| 🚨 long netted against short | NOT permitted | FAQ Q41: do not subtract shorts from longs; "report only the long position" |
| open-end mutual fund shares | NOT reported | FAQ Q7: shares of open-end investment companies "should not be reported" |
| securities on non-US exchanges | NOT reported | FAQ: shares that trade on non-United States exchanges are not reported |
| ⚠️ bonds, commodities, currencies, cash | NOT reported | INFERRED from 13f-1(c) - equity securities on the Official List only. The SEC does not state these exclusions in those words. |
✅ Confidential treatment - the positions that are not there
Requests are made "in accordance with rule 24b-2(i) under the Exchange Act"; the public filing must show "that confidential information has been omitted". The FAQ places the commercial rationale under FOIA Exemption 4, which protects "trade secrets and commercial or financial information"; the requested period "may not exceed one (1) year". On denial or expiry an amendment follows "within six business days of the denial ... or the expiration", and it "must not be a restatement" - it "adds new holdings entries".
🔑 So the public table tells you that something is missing, never what. §7 measures what that is worth.
3. ✅ Measured - the age distribution, which is the point
draw_filing_lags() sets the filing lag; the mass piles against day 45, as a deadline does.
Measured back, calendar days after quarter end:
| mean | p05 | p25 | median | p75 | p95 | max |
|---|---|---|---|---|---|---|
| 42.5 | 34.6 | 39.8 | 42.5 | 44.0 | 45.0 | 173.5 |
Now the number that matters: the age of the position when its table becomes public, across 9,792 seeded disclosed positions.
| min | p05 | p25 | median | mean | p75 | p95 | p99 | max |
|---|---|---|---|---|---|---|---|---|
| 24 | 44 | 60 | 82 | 83 | 106 | 127 | 136 | 224 |
✅ Measured: 40.9% of positions are more than 90 days old when you learn of them, and 11.5% are more than 120 days old. It decomposes into 40.9 days of establishment-to-quarter-end plus 42.4 days of quarter-end-to-filing.
🔑 🚨 "45 days" is one of the two legs and it is roughly half the total. Quoting it as the staleness of 13F data understates it by a factor of about two, and it hides that the staleness is a distribution - the same filing carries positions 24 days old and positions 224 days old, and nothing in the table says which is which.
4. 🚨 Measured - what the table never contains
- ✅ 21.5% of positions in this panel are opened and closed inside one quarter, so they never appear in any 13F. Their mean holding period is 19 trading days; the ones you do see average 154.
- ✅ 20.7% of the positions you do see have already been closed by the filing date, and 52.4% are closed within a quarter of it.
🔑 13F is a quarter-end snapshot, not a trade record, so what you can copy is selected on
being slow. Any statement about "the manager's turnover" or "the manager's trades" built
from consecutive 13Fs is a statement about the slow half of the book. That is a survivorship
bias in the time dimension, and it is not in the five gates of
../../../fin-core/skills/research-integrity-guards/SKILL.md because it is domain-specific.
5. ✅ Measured - the A/B
Weight-clone the disclosed longs for 63 trading days, short the equal-weight universe, entered at the close of the key day. Seed 20260910:
| key | Sharpe | ann. return | ann. vol |
|---|---|---|---|
| quarter-end date (look-ahead) | 1.298 | 3.47% | 2.67% |
| filing date (honest) | 0.596 | 1.59% | 2.66% |
Over 24 seeds:
| mean | sd | p05 | p95 | > 0 | |
|---|---|---|---|---|---|
| quarter-end | 1.176 | 0.395 | 0.504 | 1.668 | 100% |
| filing | 0.640 | 0.305 | 0.172 | 1.129 | 100% |
| gap | 0.536 | 0.440 | -0.214 | 1.193 | 83% |
✅ Measured pooled retention: 54.4% of the look-ahead Sharpe survives honest keying.
⚠️ 🔑 Read the sd on that gap. At this alpha level the A/B on one panel does not
settle anything - it is negative in 17% of seeds. That is the honest report, and it is also
the point: a single-run 13F A/B is underpowered, so run it over managers or seeds and quote
the spread. Compare ../insider-form-4/SKILL.md, where the same A/B is decisive in every
seed because the mechanism there is a jump rather than a slow decay.
6. 🚨 Measured - a 13F clone is not the manager's portfolio
True book: 130% long 13F-visible equity, 60% short equity that FAQ Q41 says is not reported, and a 20% non-13F sleeve. The reconstruction is the long sleeve alone at 100%, because that is all the table contains. 40 managers:
| mean | sd | p05 | p95 | |
|---|---|---|---|---|
| correlation with the true book | 0.932 | 0.016 | 0.909 | 0.954 |
| tracking error, ann. | 8.5% | 0.7% | 7.6% | 9.7% |
| vol of the true book | 23.5% | 1.0% | 22.2% | 25.1% |
| vol of the reconstruction | 21.9% | 0.5% | 21.2% | 22.5% |
| market beta, true | 0.741 | 0.095 | 0.636 | 0.881 |
| market beta, reconstructed | 0.999 | 0.025 | 0.965 | 1.036 |
🚨 A correlation of 0.93 is what makes this dangerous. It reads as a noisy copy. It is not: the market beta is wrong by a third and the tracking error is 8.5% a year. The error is a systematic exposure, not noise, and it does not average away with more quarters. 🔑 A "manager's returns" series computed from 13F filings is someone else's book. Say so, or do not publish the series.
7. ✅ Measured - confidential treatment
Omit a share of positions from the public table, re-run the filing-date clone. One panel per seed, 20 seeds, so the rows differ only in what was omitted:
| omitted | top-conviction omitted | randomly omitted | s.e. | cost of selection |
|---|---|---|---|---|
| 0% | 0.628 | 0.628 | 0.073 | 0.000 |
| 5% | 0.600 | 0.614 | 0.074 | 0.013 |
| 10% | 0.540 | 0.610 | 0.070 | 0.070 |
| 20% | 0.525 | 0.599 | 0.077 | 0.074 |
🔑 The random column barely moves at any level; the top-conviction column falls monotonically. It is not how many positions are withheld, it is which - and the adversarial assumption is the working one, because a manager requests confidentiality for a position it is still building. ⚠️ Read the cost against the standard error: below about 10% it is inside the noise here. The request can run a year, so the amendment that finally reveals the position arrives after the information in it is spent.
8. Traps
- 🚨 Keying on the quarter-end date. §5. Use the filing's own acceptance timestamp.
- 🚨 Quoting "45 days" as the staleness. §3: the median position is 82 days old and the spread runs 24 to 224. Report the distribution.
- 🚨 Calling a long-only reconstruction the manager's portfolio. §6.
- 🚨 Inferring trades by differencing consecutive 13Fs. The difference is
held at Q_endminusheld at Q-1_end. Everything opened and closed between them is invisible (§4, 21.5% here), and a "new position" may have been opened 91 days before you see it and closed before you can act (20.7% here). - 🚨 Treating an amendment as a correction to the original. A confidential-treatment amendment "must not be a restatement" - it adds entries. Merging it back onto the original filing date rewrites history with information that was withheld at the time.
- 🚨 Building the manager universe from managers that file today. A fund that closed
stopped filing. The list of 13F filers is a survivorship-biased universe in exactly the way
../../../fin-core/skills/research-integrity-guards/SKILL.md§1 describes. - ⚠️ The $100M threshold has not moved since 1978, so the filer population has grown enormously in real terms and is not comparable across decades. A study spanning 1980-2026 is comparing different populations.
- 🚨 Share counts are not weights. The table reports shares and a value as of quarter end. Converting to a portfolio weight needs a price as of that quarter end and a total that excludes everything in §2 - so the weights do not sum to the manager's book.
- ⚠️ Position size is reported for the whole 13F filer, which may aggregate several funds with different mandates, and joint filings are allowed. "The manager" may be many managers.
- 🚨 CUSIP-to-ticker mapping is a dated join. Same
(identifier, date)rule as everywhere else; a current CUSIP table reintroduces survivorship.
9. Scripts and where this sits
scripts/thirteen_f.py - the transcribed deadline, threshold, coverage table and
confidential-treatment mechanism; draw_filing_lags(), lag_summary(), age_summary()
and invisibility(); make_panel(); portfolio() (with neutral=False for the raw long
book), run_ab(), ab_over_seeds(), reconstruction_error() and
confidential_treatment(). numpy + pandas, seed 20260910, about 15 s, no network, no file
writes, no scraper.
../insider-form-4/SKILL.md- the same A/B with a two-business-day deadline, where the gap is an announcement rather than decay and is decisive in every seed.../congressional-trading-disclosures/SKILL.md- the same A/B at 45 days from the transaction, and the bracketed amounts.../social-and-influencer-feeds/SKILL.md- selecting the manager or the account after seeing the returns, which is the bias this skill's universe trap shares.../../../fin-core/skills/fundamental-and-macro-data/SKILL.md- the EDGAR endpoints, the User-Agent requirement, and the period/filed/accepted distinction 13F inherits.../../../fin-core/skills/research-integrity-guards/SKILL.md§1 and §2 - the survivorship and availability gates this skill is a special case of.../../../fin-core/skills/portfolio-and-risk/SKILL.md- if you are going to report a reconstructed book's beta and tracking error, that is where the attribution belongs.../../../fin-core/skills/backtest-validation/SKILL.md- a gap of 0.536 with sd 0.440 is not a result; it is a reason to run more panels.